An elastic matching approach applied to digit recognition

The usual methods for digit recognition on handwritten mail addresses are either structural or statistical. The structural ones operate by expert rule application. When the extraction process of such rules is done by hand, these approaches are generally difficult to design. The statistical methods use degenerated image descriptions of the digit which consist mainly of feature vectors. The decision zone learning in the corresponding vectorial space allows vector clusterization. In such techniques, the system recognition result is directly related to the chosen measure of vector relevance. This choice is difficult and no completely satisfactory solution exists in this domain. In order to overcome the drawbacks of these methods, the author has evaluated an elastic matching algorithm. This approach models and quantifies the distortions undergone by the digit to be recognized as compared with the theoretical and ideal pattern. Consequently, there is neither obligation of a decision rule elaboration nor a need for feature vector design. The author describes the principles of the elastic matching algorithm and analyzes its performance when run on a reference set